Date of Award

1-29-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Xiaowei Xu

Abstract

Entity resolution (ER) plays a pivotal role across domains by enabling data integration and quality improvement. This dissertation delves into the evolving landscape of ER, introducing innovative approaches that redefine this fundamental task. The first contribution is TriBERTa, a novel representation learning model tailored for ER. TriBERTa sets new benchmarks in entity matching and demonstrates versatility across ER processes like data blocking and resolution. Empirical evaluations on diverse datasets showcase TriBERTa’s superior performance over existing representations, including from large language models. The second contribution explores the use generative language models like GPT-3.5 and Dolly 2.0 for cross-domain entity matching using a pioneering “train once, match anywhere” approach. Without domain-specific fine-tuning, GPT-3.5 and Dolly 2.0 shows promising versatility, though performance varies across datasets. Further Analyses provide insights into factors impacting cross-domain results. Together, TriBERTa and generative models signify a paradigm shift in ER, unlocking new opportunities for data integration and quality improvement. The dissertation concludes with broader implications beyond these specific techniques, proposing future directions to advance representation learning and generative models for robust, scalable ER across domains. In summary, this dissertation redefines ER through pioneering innovations in representation learning and generative language models. The synergistic potential of these approaches is demonstrated through empirical evaluations on diverse datasets. Ultimately, this work paves the path for enhanced data quality and integration across different domains by reimagining the landscape of entity resolution.

Download

Share

COinS